Integrating Heterogeneous Prediction Models in the Cloud

Abstract

As the emergence and rapid growth of cloud computing, business intelligence service providers will host platforms for model providers to share prediction models for other users to employ. Because there might be more than one prediction models built for the same prediction task, one important issue is to integrate decisions made by all relevant models rather than adopting the decision from a single model. Unfortunately, the model integration methods proposed by prior studies are developed based on one single complete training dataset. Such restriction is not tenable in the cloud environment because most of model providers may be unwilling to share their valuable and private datasets. Even if all the datasets are available, the datasets from different sources may consist of different attributes and hard to train a single model. Moreover, a user is usually unable to provide all required attributes for a testing instance due to the lack of resources or capabilities. To address this challenge, a novel model integration method is therefore necessary. In this work, we aim to provide the integrated prediction result by consulting the opinions of prediction models involving heterogeneous sets of attributes, i.e., heterogeneous models. Specifically, we propose a model integration method to deal with the models under a given level of information disclosure by adopting a corresponding measure for determining the weight of each involved model. A series of experiments are performed to demonstrate that our proposed model integration method can outperform the benchmark, i.e., the model selection method. Our experimental results suggest that the accuracy of the integrated predictions can be improved when model providers release more information about their prediction models. The generalizability and applicability of our proposed method is also demonstrated.